Quantum Outpost

Algorithm Zoo · Machine learning

Quantum Generative Adversarial Networks

Also known as: QGAN, Quantum generative models

First described: Dallaire-Demers, Killoran; Lloyd, Weedbrook, 2018

Status: Heuristic Maturity: Demonstrated Speedup: None proven

The problem

Train a parameterized quantum circuit to generate samples from a target distribution.

Adversarial training: a quantum generator G(θ) produces |ψ(θ)⟩, samples z ∼ ⟨z|ψ(θ)⟩|². A discriminator (classical or quantum) tries to distinguish real samples from G's samples. θ is updated by gradient descent (parameter-shift rule).

Best classical

Diffusion models, normalizing flows, classical GANs — extremely strong on real data.

Quantum complexity

O(L·N) gates per generator call for L-layer N-qubit ansatz; many calls per training step.

Our verdict

Research vehicle, not a generative-model competitor. The classical baseline keeps getting stronger faster than the quantum hardware. Best near-term use is as a distribution-loading primitive for downstream quantum algorithms — and even there, classical methods often suffice.

When to use it

When not to use it

Classical baseline

Diffusion models (Ho et al. 2020) and consistency models (Song et al. 2023) dominate image generation. For tabular data, normalizing flows and copula-based methods are highly tuned. No QGAN result has matched a strong classical baseline on natural data.

Hardware cost

Each training step requires O(L·N) gates per shot, ~10^3 shots per gradient estimate, ~10^3 steps to converge. NISQ noise compounds across this.

Key papers

Deep-dive tutorials

Last verified: 2026-05-24

Weekly dispatch

Quantum, for people who already code.

One serious tutorial per week, plus the industry moves that actually matter. No hype, no hand-waving.

Free. Unsubscribe anytime. We will never sell your email.